Manifesto

Robots are moving from labs into the real world: warehouses, homes, farms, cities, but there's a bottleneck nobody talks about.

Software ships fast because every change is tested automatically, push code, run tests, and know what broke in seconds. Decades of tooling make this possible, from CI/CD and test suites to monitoring and observability.

Robots have none of this.

A robot does something it shouldn't. So you SSH in, pull the logs, download the video, scrub through footage, and try to make sense of raw sensor streams, all to figure out what happened. One run at a time, hundreds of runs a week, with no patterns detected and no memory between sessions. Every session starts from scratch.

This is the iteration loop for every robotics team on the planet, and it's slow.

Training models isn't the bottleneck; GPUs are fast. Testing them is hard, you either build real environments or simulate them, and simulation is the likely path forward. But that's a different problem. Even after the robot runs thousands of times, in the real world or in sim, you still have to understand what happened, across every test and every deployment. That is the bottleneck: making sense of the data, at scale.

Every run generates it: camera, LiDAR, IMU, encoders, force, joint states, telemetry. But knowing what to capture is hard, and what does get recorded arrives at different rates, in different formats, unaligned and unsearchable. So it goes unused. The most complete record of what a robot did is the part no one can open.

The pieces now exist. Vision-language models can watch a run and evaluate it. Sensor analysis can surface what video can't: force spikes, control instability, drift. Agentic systems can reason over both, at scale, to reveal the patterns buried in unstructured data that no human could sit through. What they lack is context: the synced, searchable, multimodal record underneath. That record is the unsolved part.

We believe this is a solvable problem.

We're building the system that understands what robots do in the physical world. We capture what's worth recording, sync every modality onto one timeline, make every run searchable, and feed that context to the agents and engineers doing the work, so the raw data becomes something they can finally query.

No engineer can watch hundreds of hours of multimodal data and hold the pattern in their head. An agentic system can. And unlike a debugging session that ends when you close the laptop, it remembers: every run adds to what a team knows about its own robots.

This is a hard problem. Understanding physical behavior from raw data, across robot types, sensor modalities, and environments, is unsolved. We're not afraid of that. Hard problems worth solving are the only ones that matter.

If we succeed, robots iterate as fast as software. And when robots iterate fast, they deploy everywhere and diffuse into every part of society.